In this paper we propose to enhance the training data of boosting-based object detection frameworks by the use of principal component analysis (PCA). The quality of boosted classifiers highly depends on the image databases exploited in training. We observed that negative training images projected into the objects PCA space are often far away from the object class. This broad boundary between the object classes in training can yield to a high classification error of the boosted classifier in the testing phase. We show that transforming the negative training database close to the positive object class can increase the detection performance. In experiments on face detection and the analysis of microscopic cell images, our method decreases the amount of false positives while maintaining a high detection rate. We implemented our approach in a Viola &; Jones object detection framework using AdaBoost to combine Haar-like features. But as a preprocessing step our method can easily be integrated in all boosting-based frameworks without additional overhead.
展开▼
